fix(task): use acall for output conversion in async paths

[greysonlalonde]

Fixes #5230

---

Note

Medium Risk
Touches task execution/output conversion paths and adds new async conversion code; mistakes could cause incorrect structured outputs or unexpected blocking/behavior differences under async workloads.

Overview
Fixes async task output conversion to be natively non-blocking: async task execution and async guardrail retries now export structured outputs via new Task._aexport_output() instead of the sync _export_output().

Adds an async conversion pipeline in utilities/converter.py (ato_pydantic, ato_json, async_convert_to_model and helpers) that uses llm.acall and offloads the sync-only InternalInstructor.to_json via asyncio.to_thread, plus refactors shared validation/unpacking logic.

Extends async tests to assert the async conversion path never calls blocking llm.call and that function-calling JSON conversion is threaded.

^{Reviewed by Cursor Bugbot for commit d61dae8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[iris-clawd]

Clean async conversion pipeline. Reviewed the full diff:

Good:

• _aexport_output → async_convert_to_model → ato_pydantic/ato_json — mirrors the sync dispatch exactly, just with acall instead of call.
• Shared logic refactored into _build_messages, _coerce_response_to_pydantic, and _unpack_model_output — reduces duplication between sync/async paths nicely.
• asyncio.to_thread for the function-calling JSON path (InternalInstructor.to_json) is the right call since that's sync-only.
• Tests properly assert acall is used (not call) and that the to_thread offload happens.

One nit: missing trailing newline at end of test_async_task.py (last line is mock_to_thread.assert_awaited_once_with(instructor.to_json) with no newline). Won't break anything but some linters care.

LGTM ✅